Good afternoon. Welcome to this digital health and innovation session. It's entitled AI and wearable technology and cardiology, streamlining care, reducing burden and shaping the future. My name is Angelo Viviano. I'm from New York Presbyterian Columbia and my co-chair is Aki Atiyah from Mayo Clinic. We have four speakers today. Each will speak for approximately 12 minutes and then we'll have questions and answers at the end of the session and there is a microphone for you to come up to because this session is being recorded. Okay. Thank you, Angelo. Our first speaker is Nikhil Alawalewe, going to talk to us about AI driven wearable technology and clinical benefits and challenges. Great. Good afternoon, everybody. Thank you for coming and thank you to the chairman and Heart Rhythm Society for the invitation to share our work with you. To begin with, please scan the QR code to submit your questions. I'm going to start with a definition. In this context, wearables in the setting of arrhythmia management are consumer facing digital tools that process data that's captured using mobile sensors and they generate measures of both behavior and physiological function. We've got a number of different devices available on the market but really the ones we come across most in clinical arrhythmia management are the risk based devices and perhaps that's because a lot of the parameters that they collect are the ones we're familiar with, so PPG derived hemodynamic data and ECG derived from voltage data. But we would ignore at our peril the unconventional parameters that are also collected, things like activity, different behaviors like how often you use your phone, how often you're in sunlight and even GPS location that may well have an impact on clinical care. Now these devices have largely been studied in the setting of screening of healthy individuals to detect different arrhythmias and a number of studies have been performed to detect atrial fibrillation and there is case series level reports of using it for other arrhythmia detection, all of this in relatively healthy individuals. But if we look at the rest of the arrhythmia journey for our patients, there is a relative paucity of information and data regarding the utility of wearables from diagnosis to monitoring to therapeutic intervention and further follow-up and really that's what prompted our imaginatively titled AF ablation follow-up using Apple Watch trial which was really designed to measure the effect of wearables on clinical endpoints. We enrolled 172 individuals who had symptomatic paroxysmal or persistent atrial fibrillation referred for their first AF ablation and we randomized them to either receive standard care or an Apple Watch based protocol recording daily ECGs or whenever prompted to by their device or by their symptoms. This work implicitly relied on the PPG based sensors and ECG technology that had been validated in the screening studies but that was of a very different population. They were relatively healthy participants and in our study we have a group of individuals with a relatively high pre-test probability of arrhythmia, their probability of having atrial fibrillation or other arrhythmias was much higher than the cohort study previously. And we know that in these groups of patients when you use PPG based sensors to detect heart rate when the individual is sat down at rest and calm, you get pretty good correlation with gold standard measures as shown by these Bland-Altman plots here. But as soon as you start introducing movement or activity into that individual, the accuracy goes down and so do we have enough accuracy to deliver good clinical care with these devices? A team also from the UK from Birmingham in the Midlands used wearables to deliver titration of beta blockers and digoxin as part of the RATE-AF trial for individuals with permanent atrial fibrillation and they demonstrated that you can with good clinical efficacy and safety deliver care in that setting using wearable devices and similar outcomes were seen in both arms of that trial. We had to also validate the ECG labeling algorithm in these devices because again it's been designed for relatively healthy individuals and we found that in our cohort of patients with high pre-test probability that when the watch delivers a label of sinus rhythm or atrial fibrillation, it's actually pretty good. It's got a high PPV, sorry, positive predictive value of 95 to 96 percent. However the challenge in this circumstance was a significant proportion, about a third of those ECGs were inconclusively labeled. So uninterpretable by the watch but our expert over reading was able to make a diagnosis. So how we handle workflows with these ECGs in these patients still needs to be determined and I think this is a good time to launch a poll but I don't know how we can answer so you can perhaps think to yourself what is the USP or unique selling point of wearable ECGs and I think there are a number of advantages. One is patient owned and so the monitoring is patient led according to their schedule. Secondly it's very long term in the data collection, really limited by only the lifespan of that device. It also collects temporally annotated metadata so each ECG is connected to lots of other data points. Actually the cost per ECG over the lifetime of that device is relatively low and there's also that automated classification algorithm and each of these should be considered when you're interpreting data or making any recommendation for an individual to use one of these devices but to us the one that was most striking was the fact that each ECG was not being interpreted by us in isolation. Because of the method of collection each ECG was time stamped by the device and connected to 60 other data points and so we were able to interpret each ECG in the clinical context of many other data points for that individual at that time and so this could be things like step count giving a reflection of their activity, the number of workouts they'd done that week, it could be their respiratory rate, blood pressure or even their GPS location and perhaps there was a correlation between where they were and what rhythm they were getting and so this was all things which we had you are able to access using wearable ECG devices. And we were able to provide patients with this sort of clinical context to their ECG data and again if you take a person who's having an AF ablation as indicated by that purple diamond you can show that individual on the y-axis here their activity beforehand and overlay with it the rhythm recorded on that ECG so green here being sinus rhythm, red here being atrial fibrillation and then you can follow them up and look at both parameters activity and ECG and help make decisions regarding further care like further ablation treatments and what their long-term outcome was going to look like and this y-axis could very easily be swapped to number of workouts that individual did, subjective survey results of quality of life and so you can really build up a picture for that individual that you can't do with traditional ECG technologies. And we were interested in this idea of using activity as an endpoint because you can show that individual sat across from you in clinic this is what your activity was like beforehand in red and this is what your activity is like after a blanking period now that you're back here with us and is there a significant difference is there not and that allowed us to personalize that can allow you to personalize care to some extent. We found that there was an association between the objective activity scores and subjective AFEQT scores reported by our participants and on an overall level if you just looked at it from a distance there was a significant increase in activity but again going back to that idea of personalization of care this was driven by about half of our patients who had a significant increase in their activity whereas those were not seen uniformly across the board. Some patients that was because of AF recurrence and again you can see differences between activity on days when they had AF recurrence compared to normal sinus rhythm days but again the effect of activity on AF recurrence was different for different patients. Some had that first recurrence on the days they were significantly more active some when they were less so and so really this left us thinking that you can you know catheter ablation and AF recurrence do perhaps affect activity but really the interpretation of this data should be personalized to the person sitting in front of you and I guess the next question is does all this metadata and all this additional data give us anything new. We learn something that we can't get from existing technologies and so we went back to the challenges that our patients were facing and we would receive ECGs from some of our patients showing sinus rhythm when they felt great but also ECGs of sinus rhythm when they didn't feel so good and they felt like their symptoms were back and then acutely afterwards they developed atrial fibrillation and we wondered was there any significant differences that we could have detected between that yellow ECG immediately before a recurrence versus that control one and were there any differences in that metadata in that individual's activity heart rate heart rate variability that we could have perhaps used to predict that there was something coming up and so we looked at this in a sort of case control methodology where we defined a case ECG as the one in sinus rhythm immediately preceding a recurrence and the control is one that was temporarily disconnected from any atrial fibrillation show sinus rhythm with no AF in the week before or week after it and from the metadata that I spoke about we were able to build a model that was able to predict sorry that was able to accurately classify case ECGs and control ones with a reasonable amount of accuracy with a AURC ROC of 0.86 here and on mixed effect modeling these additional parameters that were measured were some of them were significantly different between the two groups and so the possibility of being able to perform preemptive interventions based on this data is something which would be an attractive prospect to study in the future. Now whenever you have this much data in your hands one of the big challenges from a research perspective is the amount of missing data that you have and on this three dimensional bar chart we're on the sorry on the x-axis you've got days post ablation you've got each individual participants frequency of ECGs per day going across the screen you can see that that initial surge of ECGs at the start but was then followed by some areas of gray suggesting the use of the device and engagement with these daily ECGs did taper off towards the end. Why that was we're unsure and so we investigated whether there was any association between the rhythm recorded on the ECG the day before they stopped recording and couldn't find any significant association there suggesting perhaps it is more random than we originally did think. And a second challenge that I want to flag is that if we are using this on an end of one basis for the individual sat in front of us it's all well and good but if we do want to try and integrate these technologies in a structured systematic way so it can be used on a wider scale then we do need to also build the platforms and algorithms to do this which we don't have at the moment but colleagues at the Mayo Clinic reported that they have integrated ECG from wearables into their system so it'd be interesting to hear their experience of this as well but you know there would be a good opportunity there to think about how we could perhaps retrain algorithms to improve that inconclusive third of one third of ECGs here to perhaps deliver better labels for our cohort of patients and you know something that the implantable cardiac monitoring community have done very well to improve their number of false positive alerts. And I just want to finish on a final more optimistic note that actually using these devices is an opportunity for us to think about how we deliver care and actually if you were to view this as just a wearable ECG or a wearable heart rate monitor in isolation we're underestimating the potential impact of these tools on clinical care and really thinking about the unknown impact of all these different data types being measured and delivered together is something that we should continue to study and actually if we are designing further trials of these devices we need to incorporate harder endpoints, harder clinical endpoints that matter to our patients so that we can understand the full effect of these devices in care. I'd like to thank our team at Barts Hospital where we conducted these studies and all the patients who participated. Thank you. Thank you very much. Our second speaker is Dr. Linda Johnson from Lund University. She will speak about reducing administrative burden through AI and wearable data integration. Thank you. Thank you dear chairman for this introduction. Let's see if I can get some slides coming here too. So if any of you have missed it there's a lot of hype surrounding AI in health care. We are expecting that it's going to reduce costs, make our work so productive that we're no longer needed, triage our patients faster and care for them in the right setting. And so this is how things used to look. So this is the life course of a patient. He starts out happy but once in a while he gets sick and then the doctor intervenes and hopefully makes him better and happier again. And you see the doctor only has to visit the patient when he's sick in this setting. But with wearables there's a risk that there's going to be additional reasons for the doctor to get involved with a patient of course. I mean these are just a few snapshots of the different models, AI models that are built for health care use and some of them are consumer driven. For example consumer led screening for AF, consumer led risk prediction and others are models that doctors themselves have to act on. Diagnostic aids, risk prediction, treatment decision supports and AI assisted planning. And so the fear is maybe that we're going to be slaves to our AI models and have to use them at the peril of our patients instead of spending time with them. So I was asked to speak about wearables and so I thought I'd give a little bit of snapshot of how this might play out for wearables. This is just a summary of the three large smartwatch studies. Remember these were rather young populations. They had an average notification rate for a regular rhythm of 1.5% and that led to either a clinical workup or seven day patches and the average detection rate was 0.06% for AF. And so that's 25 notifications per AF patient detected. It's estimated that there's about half a billion smartwatches in use worldwide. That would lead to around 8 million notifications and this is at a time when we have about 15 million too few health care worker shortages. Now a lot of these as you've already heard have an integrated AI algorithm and an ECG recording device. This is a pretty good study from Switzerland where they compared the AI in the devices to a physician and the physicians could diagnose 99% of the ECGs recorded with the devices. So there's no problem with signal quality. But the AI finds 20% of the ECGs to be inconclusive and on the other ones they have an average sensitivity of around 80% specificity of 75%. So even the ECG recording devices they are going to help us diagnose the patients to some degree but they're also going to lead to a lot of inconclusive findings. And now I'm going to present a study that we did trying to of course there could also be other arrhythmias. These are only AF detection models. So we spent a bit of time wondering what would be good enough to have an AI only analysis where we reduce we remove a human from the loop entirely and that would be the ECG technician. And before we started building that model we thought about what would be a good enough model. It has to be perfect in sensitivity. It can't miss anything because I think we have we would not accept a lot of AI misses. Fewer than technician misses human errors we accept in health care. At the same time it doesn't have to classify everything correctly because the physician will be there to fix the classification if it makes an error and it can't produce too many false positives. And so we designed the dry martini study to evaluate this AI that we built the deep rhythm AI. It you feed it lead two and three raw ECG signal. It identifies all of the QRS complex and then it has a complex network architecture consisting of seven different models that tag every single beat in the signal. And then it's a very simple translation from beats to rhythm. And we trained this model using a vast data set 250 billion annotated heartbeats. I won't go into details about that. And then we felt that we needed to evaluate this with the same kind of stringency that we would use for any other medical intervention such as a drug. So it had to be a large patient population and we had to report figures in absolute numbers. So you would be able to find out you know how many patients out of a thousand am I missing if I rely on the AI instead of a technician. So we took a data set of 14000 patients that were beat to beat annotated by certified ECG technicians working in clinical practice. And then we annotated all of those signals with deep rhythm AI. And then we wrote a script that went through all of these signals and it extracted one arrhythmia per patient if it found one. And we ended up with as you can see almost twice as many AI detected episodes as technician detected episodes. So it found arrhythmia in twice as many patients. And then of course we didn't know whether these were true or false positive arrhythmias. And so we recruited as many people as we could. That was 54 divided them into 17 teams of three expert annotators and they went through all of the detected arrhythmias beat by beat and annotated them again it took. I always include this slide because sometimes there's someone who helped me with this who's in the room and and I've never even met them. And so I included as a as a gratitude and there's lots of people put in lots of time and annotating this data. And what did we find. We found that the AI model had a vastly superior sensitivity compared to technician analysis of the data. It was 98.6 percent sensitive for critical arrhythmias had a negative predictive value of 99.9 percent actually a little bit more than that. That's the same as we consider to be an acceptable threshold for sending a patient home from the E.R. with chest pain based on just one single troponin test. And it led to fewer false negatives. The technicians would miss one in 25 diagnosis whereas the A.I. only missed one in point three percent which translates to a 14 times lower relative risk of missing a diagnosis. The cost of that is more false positives. Twelve per thousand recording days for the A.I. compared to five per thousand for the technicians and the same pattern was seen across all of the different what we consider to be critical arrhythmias. So I thought I would sum up here a little bit. What is the future of A.I. and arrhythmia care. Well we definitely want to have a situation where the patient is with the doctor and the doctor is not with the A.I. devices but we are at risk of having an expansion of burden of care from A.I. that actually hasn't been robustly shown that is actually going to reduce morbidity and death in our patients. Most likely in my opinion the best A.I. models in health care are going to be the models that we barely notice that we're using. There are a number of them already. For example all of the CT scans that we that we read are read. They're presented to us after undergoing something called iterative reconstruction which sounds a lot like A.I. to me. And so we want to remove the human from the loop so that the humans that remain can be more valuable. This is from a Swedish tech pioneer. I think it's a great quote. In the world of A.I. nothing is going to be as valuable as humans and probably the work that we do is going to be more interesting because a lot of the bulk work can be taken out of the loop. That's my last slide. Well, thank you so much to the Chairs and HRS for having me here to talk to everyone about a topic near and dear to my heart, health economics. So as we've heard, and you're attending this session, the rise of wearables is undeniable. There are just so many various sensors, it's hard to keep track of what's being developed over the last couple of years. But speaking of economics, I'll focus primarily on arrhythmia detection, specifically atrial fibrillation, mainly because these algorithms are most well-developed in terms of providing direct-to-consumer diagnoses. And I know the topic is both reimbursement and economics, but we'll focus primarily on economics or the cost-effectiveness of wearables rather than specific reimbursement because codes are so dependent on your home country and the different health systems that we work in. So in the next 10 minutes, we'll just go through some lessons learned from another area of EP, primarily extended ECG monitoring, namely ILRs, patch monitors, and the cost-effectiveness impact that these extended ECG monitoring has had. We'll talk about the cost-effectiveness of wearables themselves. We'll talk about the cost of false positives, which haven't been very well characterized. And we'll also talk about why we need adequate reimbursement if we want to fully realize the potential of wearables. So now, this is a very different population than our general wearable or our general demographic who tend to wear wearables, but there are a lot of lessons that can be learned. So the drive is atrial fibrillation, as we know, silent, strongly linked to stroke. And there's a drive to detect it early to jump into that opportunistic window where we can treat people who warrant anticoagulation and potentially modify the risk factors as well to avoid the sequelae of stroke, unnecessary hospitalizations, and associated costs. And this drive to detect atrial fibrillation has created a market for extended ECG monitors, loop recorders. All the device companies have them. They're improving their algorithms and AI technology within them. And there are numerous patch monitors, extended loop recorders available. Now, here are just some trial data from two trials, Crystal AF over a decade old, comparing loop recorders to Holters, and then a more recent Per Diem trial comparing loop recorders to a 30-day patch monitor. And this is in a high-risk population, the cryptogenic stroke population looking for atrial fibrillation. And we see that the longer you look, the more atrial fibrillation you find. But why I'm focusing on these populations is because detecting atrial fibrillation is not enough to show economic benefit. We really need to be able to intervene and prevent the sequelae of atrial fibrillation to actually show economic value. And we're still trying to understand this link between device-detected or subclinical atrial fibrillation and stroke, but our evolving understanding is that stroke risk is associated with both AF arrhythmic burden, however you defined it, you know, percentage of time in atrial fibrillation or longest duration of atrial fibrillation, as well as traditional clinical CHAZ-VAS risk factors. So we can see this table from the ESC guidelines showing this interaction where it's not just arrhythmic burden, but how it interacts with clinical risk factors. And in the recently published subgroup analysis of the Artesia trial, which was looking at treating subclinical or device-detected atrial fibrillation lasting anywhere from six minutes to 24 hours, we find that the treatment benefit is really in those with the highest CHAZ-VAS stroke risk or CHAZ-VAS greater than four. Now it turns out these determinants of stroke risk are also the same important determinants of cost-effectiveness. So a little bit of primer on health economics, what are we trying to achieve here, what's our primary endpoint? Well, this is the incremental cost-effectiveness ratio, which is simply the ratio of costs, costs comparing a new strategy to something existing, standard of care, over the clinical benefit. So the benefit of a new strategy versus comparator. So lower ICERs are better, more cost-effective, and what we're trying to do is minimize the difference in the numerator. So we want a new technology to be as, you know, least expensive as possible, and trying to maximize the amount of clinical benefit or maximize the denominator to get the ICER as low as possible. So what goes into this? Well, on the cost end, the cost of the wearable, the technology, the downstream monitoring, the cost of false positives, and the downstream health resource use to respond to false positives, as well as the costs averted from preventing what everything we want to prevent, in this case the sequelae of stroke. And the clinical benefit, what goes into that? Well, the underlying stroke risk, we want to prevent stroke, as well as patient quality of life, of course. So several cost-effectiveness analyses in the wearable space, looking at atrial fibrillation detection. So this is one done in the U.S. healthcare setting. So this economic evaluation used a micro-stimulation model of 30 million individuals aged greater than 65, so already a higher risk population than the general wearable population. And they matched the characteristics of the simulated cohort based on the age distribution, comorbidity distribution of the U.S. population. And they compared multiple different screening strategies in terms of how you apply home or wearable ECG monitors or PPG monitors, but we'll focus on this one strategy here, which seems reasonable for clinical practice, where you have someone coming into clinic, they may get a 12-week ECG, you diagnose AFib, or you don't. If you don't, they get a wearable monitor, Apple Watch, Fitbit, whatever PPG monitor that you'd like. If the PPG monitor detects or suspects atrial fibrillation, you have some type of ECG confirmation. And the way that this cost-effectiveness strategy was designed, they would receive a patch monitor. And then you could either diagnose AFib or not, and then the model would figure out what are the sequelae of, or the health benefits accrued over a patient's lifetime if you screen and diagnose atrial fibrillation. So once again, slightly higher risk population, aged greater than 65, done from the U.S. perspective and done over a patient's lifetime. And what we see is that wearables are slightly more expensive, but you actually improve the life expectancy, quality-adjusted or quality-adjusted life years, QALYs, when you use the wearable screening strategy compared to a single 12-week ECG. And the ICER is $58,000 per QALY gained, which is actually a good economic value in the U.S. healthcare system. Now going back to the determinants of cost-effectiveness, well, it really comes down to the underlying stroke risk. If you screen a bunch of 30-year-olds where the prevalence of atrial fibrillation is so low, and even if they do have atrial fibrillation, the stroke risk is so low, you're not actually going to get very much health benefit by this strategy. So we probably can't extrapolate these results to necessarily other screening populations. And then a limitation of a lot of the economic analyses in the wearable space is that most models don't fully account for the impact of false positives. They may account for it saying that, all right, there's a false positive, you may see a physician or healthcare provider for an extra clinic visit to go over things, but it doesn't account for the stress to the patient. It doesn't necessarily account for unexpected emergency department visits or downstream testing to try and sort out, is this a false positive or not? So drawing the analogies back from this is the first generation of implantable loop recorders. I know that things have improved with the second, third generation AI-enabled loop recorders. But in this single center study of 1,800 loop recorders over a four-week study period, there were 1,500 transmissions, which equates to over 360 hours of device clinic staff review, which equates to $12,000 in salary costs. And the scary part is that over 50% in this population here, the false positives, sorry, there's over 50% false positive rate mainly due to premature atrial, premature ventricular contractions. Now, going to the topic of reimbursement, at least in the U.S. model, 35% of these false positives came from the same patient or same patients, which has implications during the study period, this four-week study period. Current reimbursement models, at least with Medicare, you can't actually bill within a 30-day window. So this is 35% of work burden that goes non-reimbursed. But this is a huge opportunity, though, in terms of AI to improve workflow, reduce false positives, and really harness the value of wearables in the future. So why is adequate reimbursement important? Well, one, you want to be adequately reimbursed for your time. And reimbursement, it turns out, is key to adoption in the clinical setting. Yes, you know, we're all here in this digital health session. We believe in the potential of digital devices and wearables, but our beliefs really, you know, when it comes to clinical practice, there's a gap there. And so in the survey of members from ERA, yes, you know, people are generally favorable to wearables, but they identify time constraints and reimbursement as the key barriers to adoption in their clinical practice. And three-quarters of European countries reported no reimbursement models at the time when the survey was taken. In drawing an analogy from another area of EP, remote monitoring of pacemakers and defibrillators, we have lots of studies showing that this is beneficial from a patient satisfaction standpoint, beneficial from a clinical standpoint in terms of improving outcomes, as well as being cost savings. But despite that, global adoption of remote monitoring isn't quite there yet and largely is due to inadequate reimbursement. And in this other survey from ERA, we can see the patchwork of reimbursement where Gray really has no reimbursement for remote monitoring. So I guess just in summary here, wearables, big opportunity to improve early detection and provide economic, as well as clinical benefit to health systems, patients, and providers. Key determinants of cost effectiveness, the underlying stroke risk, who are we actually screening, monitoring costs, as well as the impact of false positives in downstream health resource use. And to really fully harness the potential wearables, adequate reimbursement is critical to improve uptake in clinical practice. And AA models will be, if they aren't already, critical to this improved workflow process in minimizing false positive rates. Thanks. Thank you very much. Our final speaker is Dr. Paul Friedman from Mayo Clinic. The title of his talk is The Future of Predictive Diagnostics with AI-Powered Wearables. Thank you, Chairman, colleagues. It's a great pleasure to be here. As I look at this slide, I think my wife is right. I might be working too much. Great. So I thought I'd start with a case. Because at the end of the day, what we're trying to do is improve the human condition. And this is a story of a 73-year-old woman with no cardiovascular disease, breast cancer seven years ago, treated with six cycles of anthracycline. She did well until recently when she developed shortness of breath, swelling of her legs, no fevers or cough. Here's her chest X-ray. And all of us think, what's the mechanism? Could we have detected this earlier? Can we treat or prevent the condition? And I'm going to come back to this case. Because when we think about it, historically, you'd go see a doctor in a clinic, episodically, and then if things didn't look good, you'd end up in the hospital. Or perhaps in the meantime, however, there's life at home. That is what we all call life, what we're doing with our time, our goals, our dreams, our aspirations. And perhaps it's peppered with lab visits or clinic visits and email and portal, but these are all discrete ice pick views. And the hope is something that could non-obtrusively continue to monitor health and provide early detection, but there are important big questions that we've already heard touched on. That is, continuous data may allow early disease detection, but it has to be seamless because otherwise people just won't keep doing it. There can't be any acquisition resistance. Who owns the data? There are serious privacy concerns. Who's responsible if there's a finding? There are expenses. What can it be used for? And then what about the labor it takes to acquire, read and interpret the data? We heard a lot about that. And then most importantly, there needs to be a demonstration of outcome or benefit. So the first question I always ask is why wearable? The signal quality isn't as good and there's a big burden of monitoring. So we really need to think about screening versus monitoring that is how often do you need to check for a condition or its change? And so typically we think of screening as asymptomatic populations. Maybe they're getting ready to embark on active, you know, something, a new sport or activity. Whereas monitoring is someone who is at high risk group or has a known condition. We wanna track the status. And it may make sense to do it frequently, either at home over prolonged intervals or perhaps in hospitals, which we already do more intensely, but with the goal of enabling effective treatment. So the key questions are, is there a clinical need, a compelling use case? If we're talking about ECGs, do the algorithms exist and continuous or intermittent? Because I divide the monitors into these broad groups, long-term but intermittent, continuous, but if something you wear on your skin, typically after a month or so, it'll just be too uncomfortable, there'll be a rash, versus perhaps continuous wearable long-term. These are all in prototype, and I'll show a little bit about this more in a minute. So I would say there are some conditions for which meet the following criteria. There are evidence-based treatments. So if we know it's there, we have a lot of data saying treating it helps, and they are often under-detected. And I put together my list. It's not complete. And in the interest of time, I won't cover all of these. Going over this is kind of like going to the zoo. You can quickly walk by and look at all the animals, or you can spend hours watching the behavior of the chimpanzees, and I'll do sort of a blend of these. So if we go back to our patient, she had been treated with anthracyclines, and this is actually her EF by echocardiogram plotted here, and this is her AIECG score for a low EF. Notice, first of all, that this was seven years after chemotherapy. You're not gonna be wearing a patch that long, right? It has to be something you can tolerate. Also notice that the score went up with the EF still being normal, providing a two-year warning. Now the question is, what would you do here? And that's an unanswered question, but perhaps you'd monitor a little more carefully, or you might consider some pharmacologic changes. And so what ultimately happens, though, is she presents, and here's the clinical diagnosis, and then you can see the ejection fraction comes down, the score is high, and the ejection fraction gradually improves, and we see, as we do with any real-world test, some chatter in the values, but they're not completely back down to normal. This other patient had no screening, no warning, presented with clinical symptoms, then is treated and has full recovery of EF and normalization of the value. Now this, I think we're all fairly familiar with at this stage, and we use a convolutional neural network, a form of neural network that models mammalian visual cortex, where the dendrites and their axons are replaced with a math equation, and importantly, it's a nonlinear math equation because biological processes are nonlinear. And this was so-called supervised learning, where we fed in an ECG, the digital voltage-time waveform, and gave it what the right answer was with an echo. So we'd feed in an ECG and ask the network, what's the ejection fraction? It would have no idea. It's a computer, and it's a bunch of numbers, and it would guess 35%. And we'd say, no, this one is 50%. And through a process of back propagation, it starts changing the weights and biases, and these math equations are all slightly adjusted to minimize an error function. And what's happening, in essence, is the data are training the network. It's learning from the data, so that it's identifying multiple nonlinear patterns that humans can't effectively detect. And when we did that, and we looked to see if we could find heart failure, in fact, it was a very powerful tool. We often use the area under the receiver-operator characteristic, and we know that for a treadmill test, it's 0.85. BNP for heart failure, 0.7 to 0.8. And this test, the computer's ability to find a weak heart pump, 0.93. And just to make it real, here's a case, a separate case from a 12-lead, and I'll get to the wearables in a second, a 35-year-old man with no symptoms who comes to medical attention because his sister tragically and unexpectedly collapses. Look at the ECG, and in your own mind, what would you say is going on here? The standard computer calls it normal. AI reads it and says there's a high probability of ventricular dysfunction. The echocardiogram says the F is 18%, and in fact, there was a familial dilated cardiomyopathy, or a titan mutation. We've since demonstrated in the hands of primary care clinicians that it could be rapidly deployed because it's software-based, that when we randomized clinicians to those that had the tool and didn't, those who had it increased the new diagnosis of this life-threatening diagnosis by 30%, and that it was effective. But then what about wearables? Well, it's interesting, but that wearables acquire data in noisy, real-world environments. When you're getting an ECG, you're lying supine. A technician prepares your skin, tells you to hold still. If it doesn't look good, they repeat it. A wearable ECG, you might be slouching on your sofa after dinner and just tapping on your Apple Watch. I wonder what's going on here. And so body position, ambient noise, physical motion are all impacted, so they're filtered a lot. Here's a lead-one ECG from a patient who you can see has a CRT device. You can see the RV and LV impulses and the waveform. And this is what it looks like on an Apple Watch, which is lead-one. You don't even see the pacing pulses. You can see some changes here. So we had to retrain the network to learn to read the Apple Watch ECG. And in my sort of simple-minded analogy that I've often used, it's like it already knows how to read English, but it was never taught cursive. And so you have to say, well, here's what cursive looks like. And so it can pretty quickly adapt. So there was a form of retraining the network. We built an app. And as we've seen with many of these, because they're software-based and remote, we could enroll patients remotely. And one part-time study coordinator in a span of six months enrolled 2,500 patients from 46 states and 11 countries and obtained 125,000 ECGs. So a ton of data. And we found the patients were compliant. That is, perhaps because they knew the data was linked to a medical center. Even though they wouldn't be called with abnormalities, they could call their clinician. And with that, there was consistent engagement throughout the study, often matching the push notifications. And the presence, as was noted earlier, of a dashboard where, in this case, a patient with atrial fibrillation told me she was feeling well on one date. I called her on the phone, or rather she called me and we're talking on the phone. And so I looked up the dashboard and I could see she had developed atrial fibrillation that correlated with symptoms and then reviewed the tracings. And so essentially, when we analyzed the ability of those tracings to find ventricular dysfunction, we found that it remained a powerful test. From the watch, the AUC was 0.88, suggesting that we can use these devices, but we have to recognize integrating it into workflows, identifying when to do something are difficult questions. Other things that we want to treat. If we're gonna treat someone, for example, if they're having symptoms and we wanted to say take extra LASIK at home, it might be useful to know, for example, what their potassium is. And in the interest of time, I wasn't gonna go into this, but just underscore that not surprising to any of us, if you have an ECG with AI, you should be able to identify whether or not there's hyper or hypokalemia, which is critical before we give many of the treatments we recommend to patients. But the other group I wanna touch on before I go to the really, what I'll call newer wearable options, is this group which I think has been underrepresented in some ways. In the United States alone, there are 3.7 million births per year, and that was in 2019, 140 million globally. And we've done some work using three electrodes on the chest with a stethoscope, but it could obviously be done in a wearable form factor, showing that in a small pilot study in the US, 6% of patients in this high-risk group had peripartum cardiomyopathy. Think about that for a minute. Two vulnerable patients, a young woman and her baby, or unborn child. 140 million globally, even if these numbers are an overestimate, think about the tens of millions of people who are at risk for a condition for which there are treatable therapies. And in fact, my colleague Demi Adedensewa recently did a study with colleagues in her native Nigeria, showing that by using these tools, now let me make, I'm gonna go back one to show one other point. Look at the performance of this test, by the way. It's almost perfect. And if you think about it, the AI ECG algorithm in a primary care adult clinic has to differentiate between an older person with hypertension, diabetes, and other comorbidities that can cause some ECG abnormalities and ventricular dysfunction. Here, the classification problem is young, healthy woman, young, healthy woman with depressed ventricular function. So it seems to be an easier problem for classification. And again, she showed in this paper published last year that prospectively, in multiple clinics in Nigeria, the device works well. Here's another case for you to look at the ECG. 25-year-old woman with some mild, somewhat vague symptoms. Do you think she has heart disease? Pretty unremarkable ECG. Not surprisingly, hypertrophic cardiomyopathy is very powerfully detected by these algorithms. And in fact, her probability was very high. She had an obstructive hypertrophic cardiomyopathy, undergoes myectomy. Now, all of us would look at her ECG and go, boy, that looks abnormal. But notice that the AI says the probability is way down. And what's interesting then, you say, well, look at that. The surgery lowers the AI hypertrophic cardiomyopathy risk. Would medications do it? Turns out they do. Here's a study done together with Jeff Tyson in San Francisco from the Pioneer OLA study of Mevacampton. And notice that we have AI ECG scores at baseline and on Mevacampton therapy, and the score goes down. And that drop in score correlates with LVOT gradient and NT-proBNP, suggesting that, in fact, maybe there are some cohorts, if we pick them right, where this kind of monitoring could be useful. Now, there are a number of new tools, ECG shirts, ECG and phonocardiography radar shirts, but the one that's kind of captured my imagination is this wearable echocardiogram. Because if you really wanted to instantly monitor someone while I know their ejection fraction, that'd be an interesting way to do it. But you have to say, people move. Are you gonna get bad images? Will the images slide around? Well, let me show you some work that was really spearheaded by Dr. Atiyah. Take a look at this image. What do you think the ejection fraction is? It's like an impossible question. It's a still frame of an echocardiogram, right? Ejection fraction is squeeze. But if we show that to AI, it says, oh, EF is 64%. And then if we show the video and have humans look at it, you go, oh, it could be right. And in fact, my colleagues who are echocardiographers said it was 68%. When we showed AI the whole tracing, it said it was 70%. So that's just the first pass. We might say, what's the sex of this patient? Anybody here tell from looking at it? AI says it's a woman, and she is. And then we say, what about her age? And again, AI can tell that. So it has a lot of information. Our bodies encode a ton of unique, specific physiologic information that these signals, whether it's the signals we give off spontaneously like ECGs or backscattered energy, ultrasound, CT, MRI, can encode and embed in ways that are subtle, nonlinear, and multiple that's hard for us to read. So in conclusion, technology today is powerful enough to enable screening and monitoring of a number of significant conditions. And it's really poised to enable monitoring of these outside of the hospital and to accelerate evidence-based therapies. But there are important gaps. Proof of improved meaningful outcomes of frequent monitoring are lacking for most conditions. Reimbursement is concomitantly limited. Data ownership and privacy needs to be clarified. But medicine is being transformed. And I think we as clinicians have to be proactive rather than reactive. And just to plug, we have a core lab at Mayo that we're eager to partner for any academic centers who wanna do joint research together. And with that, I'll stop. And thank you for your attention. Thank you. Thank you very much. So this, we've got about 10 minutes for questions. There are two ways to do this. One, you can scan the QR code and submit your question and we'll read it out. And the other is to come up to the microphone. And any questions from the audience at the mic? So the first question we had, I believe it's to our second speaker, is, is the software you showed available? Is it open source? Can people use it to reduce their burden? It is not, sorry, it is not open source code, it's a commercial product, which is in use, but it's commercially developed. I have another question for, oh, sorry, we have a first, yeah. Thank you, all the speakers, for a really nice session. It's really nice to see the improvement of, like, the wearables and the AI technology and everything. But in terms of wearables, it was really nice to see the economic burden and things like this. So do we have a consensus on what the new care strategy should be for the wearables? And then, I guess we need a plan to take the right approach with, like, what was, I forget the name of the economic burden that was mentioned by Dr. Chu. But do we have a consensus on what the right strategy should be and how much time should we monitor these patients under which kind of program and take it to the insurance providers or the CMS? That's a great question. I think a lot more work can be done. As you know, the field is so rapidly evolving in terms of what wearables are available. And it's not just the hardware, but the software keeps evolving so quickly. With any economic analysis, I kept harping on the potential for false positives. But I guess drawing an analogy to kind of the loop recorder realm, like with the second generation, their false positive rates and their need for processing time has been cut by more than three quarters. And that all plays into kind of the economics behind that. So I think it remains to be seen right now. I do think we do need some additional, better, more comprehensive economic analyses to fully understand the reimbursement pathways. The last big comprehensive one was done by the UK Health Technology Group, NICE. But that was done several years ago, and they found that wearables for screening for AFib wasn't cost effective at the time. So I think now, or at least in the upcoming year, it's due for another relook. All right. Thank you. I'd like to just make one comment. That's a great question. I would say it's a call to action for the companies that are now selling these AI-enabled tools. That is, identify cohorts, pregnant women, people who've been treated with chemo agents, patients who've had a stroke to screen for atrial fibrillation. Those are the ones where we know we have evidence-based therapies, but we haven't done the trials yet to demonstrate. So I think between government funding organizations and industry interested in this area, those trials should be done. Because then when you demonstrate you are impacting lives, preventing death, preventing stroke, then these discussions go away. There's an evidence base to really drive it. I just really wonder, outside of the 30-day monitoring or 14-day monitoring with holters, if we provide Apple Watches to everyone, what the rate of the increased number of events that we're going to capture is going to be and how that will impact outcomes. Okay. We're going to move on to the next question. Thank you. Dr. Friedman raised a very important point in his talk with regards to clinical benefit. You partially answered my question, but my question is that do you think are we getting ahead of ourselves without hard clinical data with regards to outcomes? So the biggest example is, for example, with loop recorders, you monitor, you catch more AFib, but with the loop trial and other trials, the net clinical benefit, the question has been raised whether there was a net clinical benefit or putting everybody on anticoagulation or not. So similarly, yes, you are able to identify these things, but are these things ready for prime time the way they are being marketed and stuff? Or should we wait for hard data for outcomes before embarking upon these in general public? Yeah. It's a good question. It's a little bit broad, and I think it depends specifically on which cohorts, which populations we're applying the tools. But I'm a big believer in having data to drive our decisions and how to care for patients. So I think the loop study was a great study, but since then, and it wasn't that long ago, additional technologies become available. That is, when part of the burden is false positives, part of the question is how much can this be fully automated? And those things will change the balance, right? Because if the alerts drop way down because the AI is better and it's rapidly changing and you're really only being alerted for clinical events, then the study would likely have a different outcome. So I think it's a really fair question and one that hopefully over the next couple of years we'll get more answers to. Thank you. Next question. Hi. My question kind of leads into that, I guess, but my name is Jennifer. I'm not a physician. I'm a cardiac perfusionist, and I'm also a doctoral candidate with doing a qualitative study on the perceptions of AI, both prognostically and diagnostically in the cardiac space, finishing in September. And I did not, I was not able to make it to the beginning, so this may have already been talked about, but I was curious about what, there's many of these AI algorithms out there or in clinical trials, but the percentage of them that make it into clinical practice seems to be relatively small from the literature I'm reading and during my structured interviews that I conducted. Are you also seeing that the perception of AI is affecting the implementation of these really important topics? Can I answer that question? I have some industry experience, so I'm CSO for a Polish medtech company that built DeepRhythm software. We have two machine learning engineers and I think about 40 other engineers that build the platforms that make it so that the AI can be implemented, because the AI model itself just outputs probabilities, right? But what you need is a report that's sent to a patient, and I think that that's one reason why a lot of models that are tested don't get, don't come into clinical practice. That might not have so much to do with patient or clinician perceptions of it, but rather the fact that these aren't things that universities can do. Okay, thank you. Last question? It's a question pertaining to Dr. Friedman's last case, showing the echo. To me as a clinician, what struck me was the large atria in this patient. And so my question is, how far are we using AI to come up with a clinical diagnosis, not just the age or the ejection fraction? Again, that's a very good question, because a lot of the ECG work, for example, didn't talk about the size of the P wave or the QRS direction or vector, but rather low EF or not. And so I think it's a question of how we train the models and what we ask them to show. Especially in an echo, I don't think it's very far until we get a list of diagnoses, like an echo report from an image. That should be doable. I suspect that in a perfect world, it was like the epiphany that Dr. Johnson mentioned. These tools will hopefully improve everyone's capabilities. One of the places we're seeing, at least if we think about one of the limitations, it's the false positives, right? So one of the things we're interested in is you do an AI ECG, and if it's positive, we don't want to send them to a cardiologist. We want to do an on-the-spot point-of-care ultrasound and then have AI help both in the acquisition of the images and in their interpretation. And if that's positive, then a human gets engaged. So I think it will suggest diagnoses. I think we're a ways out from it providing the full spectrum of care. But I do think that, look, it has an encyclopedic memory, so to speak. So why not have it suggest a list of differential diagnoses? And why not have it say, here are the things I think. And then a human can look at it and go, well, maybe, maybe not. I don't see a downside to it being a good prompt, coach, sort of guide. My two cents. I'm sure others have other thoughts on that. There's a risk that we're going to be prompted about things all the time and that we're eventually going to have, you know, a doctor's version of alarm fatigue that we see on the CCU floors. There's at least that. All right, thank you. It's the top of the hour so I'd like to thank the speakers again and for your attendance and have a good rest of the afternoon.

digital health

AI in cardiology

wearable technology

arrhythmia management

self-monitoring

Apple Watch

AI-enabled integration

cost-effectiveness

predictive diagnostics

data privacy